School of Computer and Information Science

Transcription

1 School of Computer and Information Science CIS Research Placement Report Title: Data Mining Office Behavioural Information from Simple Sensors Name: Samuel J. O Malley Date: 20/11/2011 Supervisor: Dr Ross Smith Abstract This project discussed the concept of using three simple sensors to monitor the behavioural patterns of an office occupant. The goal of this research study is to capture behaviour information of the occupant without the use of invasive sensors such as cameras that do not maintain a level of privacy when installed. Our initial analysis has shown that data mining can be applied to capture re-occurring behaviours and provide real-time presence information to others that occupy the same building using ambient displays.

2 1 Introduction This project explores the concept of using simple sensors incorporated into an office environment to capture behavioural information that provides presence information to coworkers through an ambient display. Two sensors are used to capture the state information of doors when they are opened and closed, in conjunction with a pressure sensitive cushion on the desk chair which we refer to as a seat sensor. We present how these three simple sensors can be set up to log information to determine physical presence and behavioural habits by data mining the sensor information. By itself this data means nothing, but by data mining the information over a long period of time we can start to see reoccurring patterns in the occupant s behaviour. Sensors capture data from the physical world and convert the information to analogue or digital data. Electronic sensors can capture temperature, humidity, light, and sound amongst others. We consider the complexity of a sensor in terms of the data complexity. For example, a 1.3 megapixel webcam at 24 FPS generates approximately 20 gigabytes of information per hour. Although complex sensors are very powerful for capturing data, this paper shows how useful information can also be captured from simple sensors that can be used to capture everyday behavioural patterns. By capturing the state of these three sensors we can start to build a behavioural presence model using data mining. 2 Background Sensors come in many different forms and new types are still being discovered. In this paper we are interested in capturing two events; firstly the state of a door to identify when it is open or closed and secondly to identify when an office seat is occupied. Data mining provides a mechanism of discovering useful information by analysing captured data to provide a summary of the patterns that occur. The Apriori algorithm is commonly used on massive supermarket transaction databases to discover association rules, such as what products are commonly purchased together. This is called market basket analysis (Chen et al. 2005).

3 2.1 The Problem Research students commonly need to meet with their supervising professor for short periods of time informally; typically they will call first or just drop in unannounced. However, these impromptu meetings are not able to be prearranged and therefore often the supervisor will be out of the office or busy when the student arrives. In the case of the Wearable Computer Lab, our supervising professor is Bruce Thomas and he resides on the third floor while the lab is on the first floor. There is a lot of wasted time climbing all the way to the third floor just to find out that he is busy or not even present. The solution to this problem is to detect when the professor is present, and communicate this information to the research lab in a convenient way. However this presents another problem of how to detect the professor s presence without invading his privacy, also how to detect if the professor is busy or not. This project presents a non-invasive way of monitoring the physical presence of the professor using simple sensors, and using data mining techniques to generate a behavioural model which can be used to detect if the professor is busy or not. 2.2 Similar Applications Ambient displays 3 Research Concept 3.1 Simple Sensors While a camera in the office would provide us with all the information we need to monitor an occupants presence, it can be very invasive. By utilising simple sensors such as magnetic door sensors and a seat sensor, the only information recorded is that which could be obtained by walking past the office and looking in. By employing the state information of these three sensors, we can discover behavioural information such as the most likely time that the occupant is available as a pattern observed over a period of time. From analysing the layout of the office in Figure 2, we noted that there are three possible locations for simple sensors that provide the most meaningful information. The office we

4 are interested in shares a reception area with another office. As either door can be locked to outsiders, we chose to place magnetic door switches on both the inner and outer doors. Finally we decided to place a sensor on the desk chair to determine if the occupant is seated or not. These three states can give us an insight as to the occupant s availability. By using simple sensors we guarantee that the occupant s privacy is not broken. Even if a hacker gained access to the sensor data they cannot spy on the occupant. If a complex sensor was used, such as a webcam, then the hacker (or disgruntled research student) could read what the occupant is working on and have a much larger impact on their privacy. 3.2 Data Mining The goal of data mining the sensor information is to discover patterns in the office occupant s behaviour. This will give us an insight into the most likely times that the occupant is available. We used the Apriori algorithm (Agrawal & Srikant 1994) to generate a list of association rules. In order to apply the Apriori algorithm we logged the data as transactions, where a new item is added to the log when a state changes. An example of the data is shown Day 'Time of Day' 'Outer Door' 'Inner Door' Seated Monday,Afternoon,0,0,1 Monday,Evening,1,1,1 Etc This data is formatted to be read by the Weka Data Mining Software 1 using the Apriori algorithm to generate some association rules in the data. The first attribute is the day of 1

5 the week, followed by the time of day and the Boolean simple sensor data. This data is logged whenever any of the sensor states change, such as when a door is closed or the occupant sits in the chair. 3.3 Ambient Display of Information A visualisation we employed to present the status information to others in the same building but in a different room follows a traffic light metaphor (see Figure 1). The purpose of the Traffic Light visualisation is to inform the current availability of the occupant. This visualisation utilises the behavioural knowledge gained from preliminary data logging, based on four weeks of data, and attempts to determine whether the occupant is available or not using the state changes. The traffic light uses ambient lights to display and communicate information without distracting the user from their primary task. We employ the philosophy of Ishii et al. used in their ambientroom system. The ambientroom uses simple lights to convey information such as the number of hits on a website (Ishii & Ullmer 1997). Figure 1: Traffic Light Visualisation The traffic light displays Green when the occupant is available, Red when the occupant is not in the office, and Yellow when the occupant is present but not available, such as when he is in a meeting. By analysing the preliminary results from a four week study we can apply our knowledge of the occupant s behaviour to dictate the function of the

6 Traffic Light. The simple behaviour, excluding time factors, can be displayed on a comprehensive state chart (see Table 1). Table 1: Sensor State vs Traffic Light State Sensor State Inner Shut, Outer Shut, Not Seated Inner Shut, Outer Shut, Seated Inner Shut, Outer Open, Not Seated Inner Shut, Outer Open, Seated Inner Open, Outer Shut, Not Seated Inner Open, Outer Shut, Seated Inner Open, Outer Open, Not Seated Inner Open, Outer Open, Seated Red Yellow Red Yellow Red Yellow Yellow Green Traffic Light State The time of day can also have an effect on the Traffic Light visualisation. If it is Early Morning (before 9am) then we will display Red on the Traffic Light, even if the doors are opened. It was evident from the data that the doors open for about 5 minutes every weekday between 4 and 5am, we attributed this to the cleaners emptying the bins. The limitation of this visualisation is in its simplicity. We only show Red, Green and Yellow, but there is more information provided by the data which we do not show. This visualisation does not take into account the transitions, only the current state. If the state changes from Green (inner open, outer open and seated) to Inner open, outer shut and not seated, then we can be sure that the occupant is just stepping out of the office for a short period of time. From observing the running system we have seen that when this special case occurs, more often than not the professor appears in our lab within a few minutes of this transition occurring. Because of this limitation, a secondary visualisation was employed. We displayed a floor plan diagram of the office on a plasma screen mounted in the Wearable Computer Lab. This floor plan updated in real time to show the current state of the office, including both door sensors and the seat sensor state. This visualisation does not do any interpretation of the data and it is up to the viewer to determine from the display, if the occupant is present or busy (see Figure 2).

7 Figure 2: Floor Plan Visualisation An extension to this visualisation also displays the previous state and the time of the state change. This data can be used to determine how long the occupant has been away or seated and can assist in deciding if the occupant is present or not. 4 Implementation We employ magnetic door switches and a foam based pressure seat sensor for the implementation of our system. These sensors are monitored using a micro-controller (see Figure 3) that captures when the state of the sensors changes. Magnetic door sensors are used to acquire state information from two doors located in the office, the state identified is either open or closed. These sensors are very simple, the switch is located on the door frame and a magnet is placed on the door. When the door is shut the magnet pulls the switch into an open position. This information is identified by a micro-controller and communicated to the computer as a string of comma delimited integers over the serial port.

8 Figure 3: Seat Sensor Microcontroller We decided to create a cushion which features four sensors made from a material called Digital Foam to detect when the occupant was seated. Digital Foam is a conductive foam material that changes resistance when it is compressed (Smith 2009) this material has also been used in an input device for free-form 3D modelling which records the depth and location of finger presses on its surface (Smith, Thomas & Piekarski 2008). By reading the resistance values with the microcontroller, we can detect if the cushion is being compressed. Including four sensors enables us to potentially differentiate between a person and inanimate objects, such as a stack of books, placed on the chair. The cushion was constructed in the following steps: 1. Slice a seam on the foam cushion and remove the foam insert 2. Mark the foam to place the sensors symmetrically in each corner (see Figure 4) 3. Cut out the foam and insert a matching sized piece of digital foam to fill the holes (see Figure 5) 4. Attach four separate pieces of conductive material to cover the bottom of each digital foam piece, and one large piece of conductive material to cover all four sensors on the top 5. Crimp wires onto each piece of conductive material and sew cushion back together (with wires protruding from the cushion case)

9 Figure 4: Marked Foam Insert Figure 5: Digital Foam Pieces inserted In constructing the microcontroller to read the cushion values we created a shield (see Figure 3 above) which simplifies the connection to the cushion, and can also control the Upper and Lower voltage references to give the greatest sensitivity of results. Before this was attached to the cushion, it was tested with four potentiometers (see Figure 7). Digital foam works much like a potentiometer; the foam itself functions as a variable resistor, and by introducing a fixed resistor we can take a voltage reading between these and this is proportional to the resistance in the Digital Foam (see Figure 6)

10 V+ GND Read Voltage Here Digital Foam Variable Resistor Fixed Resistor Figure 6: Digital Foam Resistance Diagram Figure 7: Testing the Cushion Circuitry There are two modules to the software architecture, Data logging, and visualisation. The Data logging software reads the sensor information from the serial port and converts it into a meaningful format. This information, along with the time and date, is stored in a file which can later be used for data mining. This information is sent across the network using TCP/IP in a standard client-server configuration. The visualisation software then presents the data using a variable floor plan diagram on a wall mounted plasma screen,

11 and via the ambient Traffic Light visualisation. This diagram features animated doors and seat state changes. 5 Results The data used to generate the results was gathered over four weeks of constant logging. Each log entry contains the time of day which is: early morning (before 9am), morning (9am to midday), afternoon (until 5pm) or evening (until midnight), as well as the day of the week. In this paper we are focusing on finding the associations in the data rather than graphing behaviour over time because we have only recorded four weeks of data. In the future when we have logged months or even years of data then we will use graphs to visualise the occupant s reoccurring behavioural patterns over time. 5.1 Raw Results Table 2: Top Results from the Apriori Algorithm Rule Association Rule Confidence 1 100% 2 94% 3 72% 4 66% 5 100% 6 97% 7 92% 8 97% 9 96% 10 97% 11 89% 12 6% 13 87% 14 69% 15 91% 16 95% 17 79% Key:!inner = Inner Door Shut, inner = Inner Door Open

12 5.2 Discussion The association rules generated from the Apriori algorithm give us some useful insights into the office occupant s behaviour, Table 2 is a list of the most relevant association rules that we have selected. Rules 1 to 4 show us that even if the door is open before 9am, it is very unlikely that the occupant is seated yet, and Afternoon is the most likely time that the occupant is seated. Rules 5 to 7 can be interpreted as showing that the Afternoon is the most likely time for the occupant to be working with the office door shut. Rules 8 to 11 show 10 cases in four weeks of working with the inner door shut, 11 cases with the outer door shut, and 5 cases with both shut. This shows that it is a standard behaviour of the occupant to work with the door shut; however it is not done all the time. This means that if either door is shut but the occupant is seated, we display a yellow light on the Traffic Light to signify that the occupant is busy. Compared to a green light which shows the occupant is available, and a red light to show unavailability. Rules 12 to 16 show the days on which it is unlikely for the occupant to be seated, Monday is very likely that the occupant is seated and could possibly mean that Monday is the day where the occupant sits for the longest time without leaving the office. If we count how much activity occurs on each day then Monday and Wednesday are the most active days, with Thursday and Friday being the least active. Together these two rules would suggest that Monday is the busiest day in the office. However the scope of this system is only the activity which occurs inside the office, Thursday and Friday could be equally as busy days with many meetings and commitments outside the office. Rule 17 shows that 79% of Wednesday s activity was with the Inner Door open. This could mean that this is a day where the secretary and the other office occupant were not present, so that when the occupant left his office he locked the outer door rather than both doors.

13 6 Future Work We have plans to re-implement the current system using wireless communication and batteries rather than obtrusive and annoying cables. This will allow the system to run without getting in the way of normal activity. Also it provides a greater flexibility to the system because it becomes easier to add new sensors and move the system to another location. In the future we would like to introduce more publicly available data into the behavioural model to generate more comprehensive patterns and a more accurate estimate of the likely times the occupant will be available. This data includes important semester dates such as exam timetables, holiday times, student consultation times and conference deadlines. 7 Conclusion What we hope to achieve with the analysis is to provide a summary of the behavioural information over time. From these preliminary results we can already see glimpses of what we could potentially learn from a larger dataset. Given the system is kept running, the data mining will allow new patterns to be revealed and potentially re-occurring behaviours over many years could be identified. Providing co-workers with this information where appropriate will allow them to select the most appropriate time to find the occupant.

14 8 References Agrawal, R & Srikant, R 1994, 'Fast Algorithms for Mining Association Rules in Large Databases', Proceedings of the 20th International Conference on Very Large Data Bases, pp Chen, Y-L, Tang, K, Shen, R-J & Hu, Y-H 2005, 'Market basket analysis in a multiple store environment', Decision Support Systems, vol. 40, no. 2, pp Ishii, H & Ullmer, B 1997, 'Tangible bits: towards seamless interfaces between people, bits and atoms', Proceedings of the SIGCHI conference on Human factors in computing systems, Atlanta, Georgia, United States, pp Smith, RT, Thomas, BH & Piekarski, W 2008, 'Digital foam interaction techniques for 3D modeling', Proceedings of the 2008 ACM symposium on Virtual reality software and technology, Bordeaux, France, pp Smith, RT 2009, Digital foam: a 3D input device,